LTX I2V Inplace & Upsample

Two LTX I2V chores, one node

Image-to-video on LTX usually means doing two fiddly things to the latent after you've built it: optionally upscaling the video latent so you sample at a higher resolution, and injecting your starting image into the latent at a specific frame so the model knows what frame 0 looks like. In stock ComfyUI that's the LTXVLatentUpsampler node and the LTXVImgToVideoInplace node, wired in sequence with a preprocessing step in between.

easy ltxI2VInplaceAndUpsample folds both into one node - "inplace" is doing real work in the name, since the image guide is written straight into the latent in place. It slots naturally into the pack's LTX chain: encode → this → easy ltxSamplerSimple.

How it works

The source is short and orderly:

1. Upscale first. Every upscale_models you connect gets applied via LTXVLatentUpsampler - you can chain several latent upscale models and they run in sequence. This happens unconditionally, before anything else.
2. Then the bypass check. If bypass is true, it returns the (possibly upscaled) latent untouched - useful for keeping one graph that can run with or without the guide.
3. Then the image guide. If image is connected, it picks images[img_index] from the batch, preprocesses it with your img_compression, and calls LTXVImgToVideoInplace with your strength - writing the guide image into the latent so the first frame is anchored to it.

Note the ordering: the upscale runs even when bypass is on. So if you want upscale-without-guide, leave bypass on and keep your upscaler connected. If you want neither, bypass both and it's a pure pass-through.

The inputs that matter

• video_latent - the LTX video latent you built earlier (e.g. from easy ltxMultiTrackEncode). Required.
• vae - required; the upsampler and the inplace step both need it.
• image - the guide image (optional). Without it, the node is just an upscaler / pass-through.
• upscale_models - a LATENT_UPSCALE_MODEL, typically from an LTX latent upscaler loader. Optional.
• img_index (0) - which image in the batch becomes the guide. Out of range → a clear error, not a silent grab.
• img_compression (18) - LTX's image preprocessing compression factor. Higher = more aggressive compression of the guide; the LTX community's CRF-style sweet spot tends to sit around 30–40 when tuning quality, and the default here is a mild 18.
• strength (0.7) - how strongly the guide image holds the first frame. Lower = more freedom for the model to drift from it; higher = stricter anchor.
• bypass (false) - skip the image injection (upscaling still runs).

The single output is the modified video_latent - same latent, now upscaled and/or guide-injected, ready for the sampler.

Wiring and traps

Install is the pack install (Manager → ComfyUI-Easy-Media, or git clone + restart). No extra dependencies beyond having LTX-2 core nodes in your ComfyUI.

The ways this bites people:

• Expecting upscale to be skipped by bypass. It isn't. The bypass only skips the image injection. If you want a plain pass-through, disconnect the upscaler and flip bypass.
• img_index off the end of the batch. If your image batch has 4 frames and you ask for index 7, it raises image_index 7 is outside the available image range. Check your batch before blaming the node.
• Feeding it a non-LTX latent. This uses LTX's own upsampler and inplace logic; a Wan or MiniMax latent will misbehave or error. It says LTX in the name, and it means it.
• No image, no upscaler, no bypass - you just get your latent back unchanged. Not a bug; that's the node being politely optional about everything.

For LTX-2.3's standard I2V path, this is genuinely the "one node to rule the fiddly parts" - the pack's own example workflows (Templates → ComfyUI-Easy-Media) lean on exactly this shape.